Sections
Neuroinformatics - 2023
SESSION 1
Monday, October 23 18:30 – 19:15
Lecture-hall Актовый зал
Chair: Prof. DMITRY YUDIN
Neural network theory, neural paradigms and architectures
1. SHIBZUKHOV ZAUR MUKHADINOVICHMoscow State Pedagogical University
Some robust variants of the principal components analysis
Abstract. Two new robust variants of the formulation of the problem of searching for the principal components are considered. They are based on the application of differentiable estimates of the average value, insensitive to outliers. In principle, this approach makes it possible to overcome the influence of outliers in the task of searching for the principal components. The effectiveness of the proposed approach is clearly demonstrated on real data.
2. * DEMIDOVSKIJ A.V., KAZYULINA M.S., SALNIKOV I.G., TUGARYOV A.M., TRUTNEV A.I., PAVLOV S.V
1Higher School of Economics, Nizhny Novgorod Branch
2
Towards integrating Hebb rule into training of Convolutional Neural Networks
Training acceleration is one of the prominent research directions in the field of deep learning. Among other directions in this field, Hebbian learning is considered to be a highly prospective approach due to its locality and highly parallel nature. Moreover, there is an emerging trend of applying Hebbian learning as a part of mixed training strategies that might include various backpropagation methods. Also, Hebbian learning is plausible for neuromorphic hardware In this paper, we overview existing approaches of applying Hebb rule to training one of the largest and most demanded classes of deep neural networks - Convolutional Neural Networks. We analyze the availability of existing software solutions for Hebbian learning. More importantly, we investigate various approaches to the implementation of Hebb rule to convolutional layers as they are foundational for modern deep neural networks.
3. * VITIUGOVA IULIIA MIKHAILOVNA
Lomonosov Moscow State University
Comparison of Input Feature Selection Methods Based on Neural Network Weight Analysis
In this research, a comparative analysis of the effectiveness of various methods for analyzing the significance of input features for neural networks is conducted. It uses the example of a gas mixture composition determination task based on semiconductor sensor data. Throughout the work, evaluation and selection of input features were carried out using neural network weight analysis methods and deep Taylor decomposition, as well as comparison with reference methods.
SESSION 2
Tuesday, October 24 14:00 – 15:15
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. KARANDASHEV YAKOV
Adaptive behavior and evolutionary modelling
4. LAZOVSKAYA T.V., TARKHOV D.A., CHISTYAKOVA M.A., RAZUMOV E.M., SERGEEVA A.A., PALAMARCHUK V.`Peter the Great St. Petersburg Polytechnic University
Analysing Family of Pareto Front-Based Evolutionary Algorithms for PINNs: a Case Study of Solving the Laplace Equation with Discontinuous Boundary Conditions
An evolutionary algorithm based on the Pareto front to construct a solution to an ill-posed problem with multi-criteria is proposed. It incorporates information about the desired solution at different stages of training neural network models. The Laplace equation in the unit square with discontinuous boundary conditions is used as a case study. The algorithm is compared with the classical one, and significant advantages are demonstrated, the influence of hyperparameters on results is studied.
5. VLADIMIR B. KOTOV AND ZAREMA B. SOKHOVA
Scientific Research Institute for System Analysis, Moscow
Unawareness as a cause of determinism violation. A metaphoric model
The research models the functioning of an autonomous agent or group of such agents. The aim is to investigate the significance of unawareness as a main cause for the unpredictability of events. The model world looks like a cellular field (two-dimensional array). There are two types of subjects: agents and observers. Because of deficient perception the subjects have limited knowledge. Every agent has an energy reserve and certain information including the map of the passed way. The computer simulation is used to investigate the model. When dealing with an agent, several scenarios are possible: 1) the agent masters the working space fully, 2) the agent harnesses part of the space to meet his energy requirements, 3) the agent explores part of the space that cannot provide him with sufficient energy. To investigate the effect of the environment, we put agents in the working space of a complicated topology where cells make up two or more congruent two-dimensional arrays rather than one two-dimensional field. Additionally, we describe different mechanisms of the interaction of groups of agents and analyze possible results of such interactions. The study shows that even a closed system with deterministic laws can demonstrate unpredictability.
6. * GAVRIIL KUPRIYANOV, IGOR ISAEV, SERGEY DOLENKO
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Study of Modifications of Gender Genetic Algorithm
This study compares several modifications of a gender genetic algorithm (GGA). Aside the difference between the genders in the probability of mutation, we introduce two additional modifications: different implementations of selection and different laws of dependence of the probability of mutation on gene number within a chromosome. We use four test optimization problems in spaces of various dimensions to compare conventional GA, conventional GGA, and GGA with the additional modifications implemented separately or together. It is demonstrated that the proposed additional modifications outperform conventional GA and conventional GGA in the achieved value of the fitness function, especially in high-dimensional spaces. With the increase in the problem dimension, they degrade more slowly. Also, the new modifications prevent premature convergence of the algorithm.
7. MIKHAIL MAKAROV, IVAN SEMENOV, NATALIA PIDZHAKOVA
A logical output system for a new type of heuristic deccisions for of adaptive control of a mobile robot
The article presents theoretical information that reveals the principles of logical output of a new type of heuristic decisions synthesized inside the control system of a mobile robot. An experimental study was conducted aimed at substantiating the effectiveness of the proposed approach to solve the designated task. It is revealed that such a variant of the implementation of this procedure contributes to the improvement of the adaptive abilities of the robot when working in a dynamic environment.
8. K.A. FEDOROV, M.YU. NAZARKO, A.V. SAMSONOVICH
National Research Nuclear University (MEPhI), Moscow
CONTROLLING NEURAL NETWORK TRAINING WITH A GENETIC ALGORITHM
The work addresses the problem of integration of deep learning and genetic algorithms (GA). An approach is developed, according to which GA is directly modifying the training datasets instead of the trained network parameters. These datasets are records of agent's behavior in the environment. They are treated as genotypes within GA, while phenotypes are the trained neural networks. The architecture and hyperparameters of the neural network and its learning model remain fixed. Results of numerical experiments based on the example game paradigm "Three Cowboys" prove the concept and confirm the efficacy of the proposed approach.
POSTER SESSION 1
Tuesday, October 24 14:00 – 17:00
Lecture-hall НЛК 2 этаж
Chair: Prof. ISAEV IGOR VIKTOROVICH
Neuromorphic computing and deep learning
9. MAGAY G. I., SOROKA A. A.1National Research Nuclear University (MEPhI), Moscow
2National Research University "Higher School of Economics", Moscow
Estimating the Transfer Learning Ability of a Deep Neural Networks by Means of Representations
The basis of transfer learning methods is the ability of deep neural networks to use knowledge from one domain to learn in another domain. However, another important task is the analysis and explanation of the internal representations of deep neural networks models in the process of transfer learning. Some deep models are known to be better at transferring knowledge than others. In this research, we apply the Centered Kernel Alignment (CKA) method to analyze the internal representations of deep neural networks and propose a method to evaluate the ability of a neural network architecture to transfer knowledge based on the quantitative change in representations during the learning process. We introduce the Transfer Ability Score (TAs) measure to assess the ability of an architecture to effectively transfer learning. We test our approach using Vision Transformer (ViT-B/16) and CNN (ResNet, DenseNet) architectures in computer vision tasks in several datasets, including medical images. Our work is a contribution to the field of explainable AI and an attempt to explain the learning transfer process.
10. VLADIMIR B. KOTOV AND ZAREMA B. SOKHOVA
Scientific Research Institute for System Analysis, Moscow
The Resistor Array as a Commutator
Being necessary components of large smart systems (including the brain), commutators can be realized on the basis of a resistor array with variable resistors. The paper considers some switching (commutating) capabilities of the resistor array. A switching graph is used to describe the work of the resistor array. This sort of graph provides a visual representation of generated high-conductivity current flow channels. A two-terminal scheme is used to generate the switching graph. In the scheme a voltage is supplies to a particular couple of poles (conductors), other poles being isolated from the power sources. Changing couples of poles makes it possible to generate a series of switching graphs. We demonstrate the possibility to create an interconnection between two or more blocks connected to the appropriate poles of the array. To do this, the resistor array must have a suitable signature (resistor directions), the applied voltage must match the signature. The series we generate are defined by not only control signals, but also the prehistory of the resistor array. Given preset resistor characteristics, the competition between graph edges plays an important role in that it contributes to the thinning of the switching graph we generate.
11. ALEXANDER SBOEV, DMITRY KUNITSYN, ALEXEY SERENKO, ROMAN RYBKA, VADIM PUTROLAYNEN
1National Research Centre "Kurchatov Institute", Moscow
2The Moscow Institute of Physics and Technology (State University)
3
Towards solving classification tasks using spiking neurons with fixed weights
A layer of spiking neurons with non-trainable weights, either fixed on the base of logistic functions or drawn from a uniform random distribution, is shown to be an efficient extractor of meaningful features for their subsequent processing by a linear classifier. The output spiking rates of the proposed layer are shown to allow predicting classes by logistic regression with F1-macro scores of 94%, 96% and 97% respectively for the classification tasks of handwritten digits, Wisconsin breast cancer and Fisher’s Iris. The proposed layer could therefore serve as a feature extraction layer for classification tasks, facilitating the development of compact and efficient models for solving classification tasks.
12. TARANIN MIKHAIL, TARANIN SERGEI
Research Centre "Module" (RC Module)
Color spike video camera model
A spatiotemporal model of a color spike video camera for use in neuromorphic systems is proposed. The model imitates the work of color-opponent P-type retinal ganglion cells. The main stages of information processing that occur in the retina of primates have been implemented. These steps include multi-scale spatial filtering using color-opponent filters, imitation of microscopic eye movements, and temporal filtering. The video camera can be an alternative to DVS sensors.
13. ANTSIPEROV V.E., PAVLUKOVA E.R.
Centro-lateral threshold filtration as a method for neuromorphic data coding
A new approach to the synthesis of methods of central-lateral threshold nonlinear filtering (coding) of images is considered. The approach is motivated by known mechanisms of human perception, in particular, by the universal mechanism of lateral inhibition. A statistical parametric model of images in the form of a system of receptive fields is proposed. The model allows a simple procedure for estimating the probability of counts, which, in turn, is taken as the basis of optimal coding.
Applied neural systems
14. TARKHOV D.A., LAVYGIN D.A., SKRIPKIN O.A., ZAKIROVA M.D., LAZOVSKAYA T.V.Peter the Great St. Petersburg Polytechnic University
Optimal Control Selection for Stabilizing the Inverted Pendulum Problem Using Neural Network Method
The task of managing unstable systems is a critically important management problem, as an unstable object can pose significant danger to humans and the environment when it fails. In this paper, a neural network was trained to determine the optimal control for an unstable system, based on a comparative analysis of two control methods: the implicit Euler method and the linearisation method. This neural network identifies the optimal control based on the position of a point on the phase plane.
15. KALININA A.Y., KIREEV V.S.
1National Research Nuclear University (MEPhI), Moscow
2
VISUALIZATION OF THE CORPUS OF DOCUMENTS BY EXTRACTING ENTITIES AND RELATIONSHIPS OF THE SUBJECT AREA BASED ON THE NEURAL NETWORK MODEL OF DEEP LEARNING T5
In this work, the study and analysis of existing methods for extracting entities and relationships from natural language texts, including methods for eliminating coreference, methods for extracting entities, methods for extracting relationships, was carried out. Using the open RURED dataset, a neural network model based on T5 was trained and applied to the collected RBC news dataset. The obtained relations were visualized in the form of a knowledge graph using the tools of the Neo4j graph database.
16. KULYABINA E.V., KULYABINA T.V., MOROZOVA V.V., MOROZOV V.U.
Issues of research and selection of methods and means for ensuring accuracy and validation when creating a virtual sample of a living cell for modeling intracellular processes using neuroinformatics methods
The issues of creating means and methods for studying a virtual model of a living cell using neuroinformatics methods are considered, taking into account the creation of tools and methods for metrological support for obtaining accurate measurement results, subsequent prediction of the behavior of cell organelles, and the consequences of the impact of new viruses on a living cell. Approaches to the study of cell responses to various environmental conditions (pH value, excess/deficiency of trace elements, disruption of organelle functioning, temperature effects, etc.) will be presented; modeling the behavior of different compartments of a living cell. Methods for ensuring the metrological traceability of measurement results will be shown.
17. KUZMENKO ANDREW VLADIMIROVICH, KIREEV VASILIY S.
1National Research Nuclear University (MEPhI), Moscow
2
Classification of methods for extracting relational triples from natural language texts
he paper considers methods for solving the problem of extracting relational triples from natural language texts. The authors proposed a classification of these methods into four groups: pipeline methods, methods based on filling tables, methods based on the introduction of additional labeling of text sequences, methods based on sequence-to-sequence conversion. The results of a comparative analysis of these groups of methods are presented.
18. ANDREI BUKH, ELENA RYBALOVA, IGOR SHEPELEV, TATYANA VADIVASOVA
Saratov State University
Classification of musical intervals in a chain of FitzHugh-Nagumo oscillators
The work is devoted to the study of biologically relevant neural networks, which are called spiking neural networks, and is aimed at classifying musical intervals. Each interval corresponds to its own chain of neurons with a unique topology, which allows classify elementary audio signals using the spiking neural networks.
19. NIKOLAY FILATOV, MIKHAIL KINDULOV
1Peter the Great St. Petersburg Polytechnic University
2Sanct-Petersburg State University
Low rank adaptation for stable domain adaptation of vision transformers
Unsupervised domain adaptation plays a crucial role in semantic segmentation tasks due to the high cost of annotating data. Existing approaches often rely on large transformer models and momentum networks to stabilize and improve the self-training process. In this study, we investigate the applicability of low-rank adaptation (LoRA) to domain adaptation in computer vision. Our focus is on the unsupervised domain adaptation task of semantic segmentation, which requires adapting models from a synthetic dataset (GTA5) to a real-world dataset (City-scapes). We employ the Swin Transformer as sthe feature extractor and TransDA domain adaptation framework. Through experiments, we demonstrate that LoRA effectively stabilizes the self-training process, achieving similar training dynamics to the exponentially moving average (EMA) mechanism. Moreover, LoRA pro-vides comparable metrics to EMA under the same limited computation budget. In GTA5→Cityscapes experiments, the adaptation pipeline with LoRA achieves a mIoU of 0.515, slightly surpassing the EMA baseline's mIoU of 0.513, while al-so offering an 11% speedup in training time and video memory saving. These re-sults highlight LoRA as a promising approach for domain adaptation in computer vision, offering a viable alternative to momentum networks which also saves computational resources.
20. Е.А. BOGDANOVA, V.N. NOVOSELETSKY, K.V. SHAITAN
Lomonosov Moscow State University
Binding affinity prediction in protein-protein complexes using convolutional neural network
Binding affinity is an important characteristic of protein-protein interactions, its determination is significant for the development of a wide range of drugs and biotechnological preparations. This paper presents an algorithm based on convolutional neural network that predicts the value of the dissociation constant for protein-protein complexes from their spatial structures, as well as a method for converting this data format into a suitable one for use in neural networks.
21. PODOPRIGOROVA N.S., SAVCHENKO G.A., RABCEVICH K.R., KANEV A.I., TARASOV A.V., SHIKOHOV A.N.
1Bauman Moscow State Technical University
2
Forest Damage Segmentation Using Machine Learning Methods on Satellite Images
Automation of satellite image analysis is an important task today. The article describes an approach to forest damage recognition using machine learning methods, which allows automating the data analysis process and identifying ar-eas where damage has occurred. This significantly simplifies the work of forest-ry services and increases the effectiveness of forest resource management. The article investigates forest damage segmentation on satellite images using mod-ern deep learning methods. The data consists of pairs of multitemporal Senti-nel-2 images. Models based on U-Net, MultiresUNet, Attention U-Net, Res-Net50 U-Net, MobilNetv2 U-Net architectures were created for forest damage detection on satellite images. The proposed approach demonstrates results in identifying forest damage areas, which can be useful for assessing the ecological situation in forested areas and conducting work on preserving forest resources. The experiment results confirm that the use of U-Net is a promising direction for further improvement of satellite image analysis methods in the field of ecol-ogy and forestry. The U-Net model shows the best accuracy (Dice=0.78).
22. SHAMARINA E.A., GUSEVA A. I., KIREEV V.S.
1National Research Nuclear University (MEPhI), Moscow
2
SENTIMENT ANALYSIS BASED ON NEURAL NETWORKS-TRANSFORMERS
The article discusses the results of a comparative analysis of the use of two neural network models BERT-Base and BERT-Large to determine the tonality of text on a dataset in the form of BBC news publications for 2019-2022, containing 16169 articles in English. It is revealed that both models have a high accuracy of text classification (0.88), and the values of the loss function for them are small and amount to values in the range of 0.19-0.21.
23. MAXIM I. CHULIN, YURY V. TIUMENTSEV, RUSLAN A. ZARUBIN
Moscow Aviation Institute (National Research University)
LQR Approach to Aircraft Control Based on the Adaptive Critic Design
Motion control of modern and advanced aircraft has to be provided under conditions of considerable and diverse uncertainties in the values of their parameters and characteristics, flight regimes, and environmental influences. The aircraft control system must be able to adapt to these changes by promptly adjusting the control laws used. The tools of adaptive control theory allows us to satisfy this requirement. In this case, it is very desirable not only to provide the created system with the property of adaptivity, but also to do it in an optimal way. An effective way to implementing this kind of adaptivity concept is the approach based on machine learning methods and tools, including technologies based on reinforcement learning. One of the approaches to the synthesis of control laws for dynamical systems that are widely used at present is the LQR (Linear Quadratic Regulator) technique. A significant limitation of this approach is the lack of adaptivity in the resulting control law, which prevents it from being used under conditions of incomplete and inaccurate knowledge of the properties of the control object and the environment in which it operates. To overcome this limitation, it was proposed to modify the standard variant of LQR based on approximate dynamic programming, a special case of which is the Adaptive Critic Design (ACD). For the ACD-LQR combination, the problem of longitudinal angular motion control of a maneuverable aircraft is solved. The results obtained demonstrate the capabilities of this approach to controlling the motion of an aircraft under uncertainty conditions.
24. ANDREW YU. TIUMENTSEV, YURY V. TIUMENTSEV
Moscow Aviation Institute (National Research University)
Motion Control of Supersonic Passenger Aircraft Using Machine Learning Methods
Motion control of modern and advanced aircraft has to be provided under conditions of incomplete and inaccurate knowledge of their parameters and characteristics, possible flight regimes, and environmental influences. In addition, a variety of abnormal situations may arise during flight, in particular, equipment failures and structural damage. The control system must be able to adapt to these changes by adjusting the control laws in use. The tools of the adaptive control allows us to meet this requirement. One of the effective approaches to the implementation of adaptivity concepts is the approach based on methods and tools of neural network modeling and control. In this case, a fairly common option in solving such problems is the use of recurrent neural networks, in particular, networks of NARX and NARMAX type. However, in a number of cases, in particular for control objects with complicated dynamic properties, this approach is ineffective. As a possible alternative, it is proposed to consider deep neural networks used both for modeling of dynamical systems and for their control. The capabilities of this approach are demonstrated on the example of a real applied problem, in which the control law of longitudinal angular motion of a supersonic passenger airplane is synthesized. The results obtained allow us to evaluate the effectiveness of the proposed approach, including the case of failure situations.
25. NOVOZHILOV KONSTANTIN ALEKSEYEVICH , KOMYSHEV DANIL ALEKSEYEVICH, KOZHENIKOV SERGEY PAVLOVICH, SERGEYEV VALERY GEORGIEVICH, KOZLOV DENIS SERGEYEVICH
Udmurt State University
Diagnosis of Parkinson's disease using deep machine learning from electroencephalography data
This paper investigates the use of convolutional neural networks to detect Parkinson's disease from electroencephalography (EEG) data. Forty-seven people participated in the study. The EEG was recorded from 21 leads to which coherence analysis was applied. The model showed an average reliability of 72-79% for the different frequency bands. The maximum reliability was in the theta band, which may be caused by synchronisation of the tremor with the theta rhythm of the contralateral cerebral hemisphere.
26. MEDVEDEVA T.V., KNYAZEVA I.S., MASHARIPOV R.S., KIREEV M.V.
1Bekhtereva Human Brain Institute RAS, St.Petersburg
2Sanct-Petersburg State University
Graph Neural Networks for analysis of rs-fMRI differences in open vs closed conditions
Functional magnetic resonance imaging data (fMRI) is a noninvasive whole-brain neuroimaging method, which is widely used to explore a variety of topics. FMRI data is usually acquired during resting state in at least one of the following conditions: eyes open, eyes closed, eyes fixated on a target. There are notable condition-related dissimilarities in the neuronal activity, according to a number of studies. These studies used standard mass-univariate analysis to check functional connectivity differences in open and closed eyes conditions. Here we discuss multivariate approach based on graph neural networks (GNN). The brain systems can be represented as a graph with regions of interest as nodes and connections between them as edges. The GNN model makes it possible to classify open/closed eyes conditions using graph representation of the brain connectivity. Previously GNNs were used to identify subjects with disorders from controls. Here we considered to classify resting states based on functional connectivity data of healthy controls. After model training, we apply algorithms for explanation of model’s predictions. Interpretation of the GNN model trained on rs-fMRI data can give an insight into what edges are important for classification. The proposed model classifies eyes closed and fixated with accuracy up to 80%. Areas important for classification between these conditions include: visual networks, default mode network and frontoparietal cognitive control.
27. LEONID S CHERNYSHEV, IVAN M ANTONOV
Optimal process control system using neural network twin and delta optimization method and its testing on a synthetic data array
The problem of determining the set of values of the control parameters of the technological process that optimize its target function on a fixed time or spatial horizon (control horizon) is solved. The solution of the problem is carried out in 2 stages – the first stage is the synthesis of the digital twin of the control object using a neuro-network model (INS model). At the second stage, the author's optimization method "delta" is used for the synthesis of optimal control, which allows determining a set of values of the control parameters of the control object that optimize its target function on a given management horizon. The system is being tested on a synthetic (simulated) data array. The robustness of the obtained control to change the learning error of the INS is evaluated.
28. MARIA O. TARAN, GEORGIY I. REVUNKOV, AND YURIY E. GAPANYUK
Bauman Moscow State Technical University
Generating Generalized Abstracts Using a Hybrid Intelligent Information System for Analysis of Judicial Practice of Arbitration Courts
The article is devoted to the description of the operation of the hybrid intelligent information system for analysis of judicial practice of arbitration courts. The structure of a generalized abstract of a judicial act is considered. The essential fac-tor is that lawyers perceive a verbatim quote much better, whether it be a sentence or a whole paragraph, than a generated text; in the latter case, they have no confi-dence that the essence is reflected correctly. Existing solutions to the problem of abstract generation of a judicial act are briefly reviewed. The proposed solution is based on the hybrid intelligent information system for analysis of judicial practice of arbitration courts. The main elements of the system are the subsystem of sub-consciousness (SbS) and the subsystem of consciousness (SbC). The role of the environment is performed by the texts of judicial acts that can be submitted to the system. Depending on the number of input documents, a judicial act (one docu-ment) and judicial practice (several documents) can be distinguished. The subsys-tem of subconsciousness includes the following modules: the preprocessing module, the feature extraction module, the module of clustering and grouping of judicial acts, the paragraph classification module. The subsystem of conscious-ness includes the following modules: the module for compiling a summary of a judicial act, the judicial practice analysis module, the report generation module. The results of the experiments are described in the corresponding section.
Neural network theory, neural paradigms and architectures
29. ZHARIKOV I.N., OVCHARENKO K.A.The Moscow Institute of Physics and Technology (State University)
Study of Rescaling Mechanism Utilization in Binary Neural Networks
Single Image Super Resolution (SISR) is a common task on devices to enhance quality of visual data. Deep Convolutional Neural Networks (DCNN) have recently shown great results in this field. However, DCNN are not compatible with resource-limited devices, because they demand significant amounts of memory and computations. Binary neural networks (BNN) provide a promising approach to reduce computational complexity and speed up the inference of a model. SISR models are much more vulnerable to degradation in performance when decreasing the precision of weights, than image classification models, due to the complexity of the task that relies on dense pixel-level predictions. Hence, they suffer a significant quality drop when being binarized. The paper investigates the importance of restricting information in BNN and proposes several binary block modifications based on different rescaling mechanisms. We implement different rescaling modules into the binary block and prove them to increase model performance. We conduct broad ablation study and inspect the resulting attention maps and activation distributions to determine the importance of rescaling the binary convolution output and the residual connection. Our modifications outperform existing BNN on benchmark datasets, showing the importance of the rescaling mechanism for increasing BNN quality.
30. RODIONOV DANILA
Lomonosov Moscow State University
Rational approximation of the solution's convolution kernel of wave equation in channel by means of neural neural networks for transparent boundary conditions construction
The paper offers a construction technique for rational approximation of the solution's convolution kernel of wave equation in channel by means of neural neural networks for transparent boundary conditions construction. The analysis of the obtained results depending on the parameters of the constructed model is carried out, comparison with the solution of the module scipy.Pade is performed. Perspective realms of research and application are also declared.
SESSION 3
Tuesday, October 24 15:15 – 16:30
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. KISELEV MIKHAIL
Artificial intelligence
31. DMITRY KARPOV, VASILY KONOVALOVThe Moscow Institute of Physics and Technology (State University)
Multi-task Encoder-agnostic Transformer-based Models for Conversational Tasks
The explored encoder-agnostic transformer multi-task models yield results matching or approaching the single-task models on most tasks. If we truncate the tasks to a low size (200-2000 samples per task), multi-task models start exceeding the single-task ones. This effect might depend on the number of samples per class.
32. * A S M HUMAUN KABIR; PROF. ALEXANDER ALEXANDROVICH KHARLAMOV; ILIA MIKHAILOVICH VORONKOV
1The Moscow Institute of Physics and Technology (State University)
2
Research Methods for Fake News Detection in Bangla Text
This research work focuses on the fake news classification in Bangla language using deep neural networks and machine learning classification algorithms. Bangla language is the fifth most spoken native language in the world with approximately over 300 million native speakers and another 50 million as second language speakers. In this work, news collected from different online and print newspapers are classified in authentic and fake news class. Considering prominent natural language processing techniques, data preprocessing and performing different deep neural networks and machine learning classification algorithms, it has achieved a maximum of 73% accuracy for minor class and 99% accuracy for classifying fake news and authentic news.
33. YU.T. KAGANOV
Bauman Moscow State Technical University
ON THE QUESTION OF THE DYNAMIC THEORY OF INTELLIGENCE
The paper considers an approach based on nonlinear and symbolic dynam-ics for the analysis of cognitive processes. The formation of semantic structures as a result of self-organization processes in complex nonlinear systems is investigated. To implement the proposed approach, the theory of metagraphs and granular computations is used. The possibility of using the proposed approach for the study of cognitive processes of intelligent sys-tems and the further development of artificial intelligence systems is shown.
34. * HUZHENYU ZHANG AND DMITRY YUDIN
The Moscow Institute of Physics and Technology (State University)
Offline Deep Reinforcement Learning for Robotic Arm Control in the Maniskill Environment
Offline reinforcement learning (Offline RL) has been widely used in robot control tasks, while online reinforcement learning is often abandoned due to its high interaction cost with environment. In order to face this situation, Behavior Cloning (BC), an approach of Imitation learning (IL), is often considered a suitable choice for solving this problem, in which the agent could learn from a offline dataset. In this work, we propose a intuitive way, in which add a Proximal Policy Optimization (PPO) loss as a correction term to the BC loss. The models are trained on a static dataset with four different robotic arm control tasks given by ManiSkill and we compare with other Offline RL algorithms.
35. ANTONETS VLADIMIR ALEXANDROVICH, ANTONETS MIKHAIL AKTXANDROVICH
1N.I. Lobachevsky State University of Nizhni Novgorod
2
Quantitative estimation of the relevance of data selected for machine learning
The problem of obtaining a measurable assessment of quality of training data selected by experts for artificial intelligence systems is considered. This can be done if the data can be displayed as histograms. Then, for each particular set of histograms, the variable maximal problem for zero-sum games is solved and the weight-discriminator function is computed, which allows the quantum of relevance of each particular element from the learning dataset.
SESSION 4
Tuesday, October 24 17:00 – 19:00
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. KISELEV MIKHAIL
Neuromorphic computing and deep learning
36. MIKHAIL S. TARKOV, VICTORIA V. IVANOVA1Rzhanov Institute of Semiconductor Physics, Siberian Branch of Russian Academy of Sciences, Novosibirsk
2
Image Processing with Reservoir Neural Network
Reservoir neural network (RNN) is a powerful tool for solving complex machine learning problems. The reservoir is a recurrent part of the network having a large size and rare internal connections which are most often set ran-domly and remain fixed. The idea of the RNN is to train only part of the net-work using a simple classification/regression technique and leave most of the network (reservoir) fixed. At the same time, all RNN advantages are preserved, and the training time is significantly reduced. The work performed optimization and research of methods that improve the reservoir ability to solve problems of image classification. These methods are based on the reservoir output data transformation before they are fed to the RNN output layer. In the work, the op-timal parameters values for the methods Infomax and SpaRCe were obtained, which provide a minimum error in image classification. Using the example of image classification from the MNIST handwritten digit database, it is shown that: 1. Reservoir networks are trained much faster than convolutional net-works, although they are inferior to the latter in terms of image classification accuracy. 2. ESN (echo-state network) with principal component projector (PCA) gives more accurate results than ESN, Infomax and SpaRCe networks, but is slower.
37. * MIRON M. LEONOV, ARTEM A. SOROKA, ALEXANDER G. TROFIMOV
National Research Nuclear University (MEPhI), Moscow
Russian Language Speech Generation from Facial Video Recordings Using Variational Autoencoder
This paper describes the use of a variational autoencoder (VAE) based generative adversarial neural network (GAN) to generate Russian language speech from fa-cial video recordings. The proposed system uses the VAE to learn a low-dimensional representation of the input video frames and the corresponding speech signals. A discriminator is used for regularization during training. The model is trained on Russian-language speech and video recordings of a speaking person collected from video hosting sites. The results show that the proposed method can generate high quality speech signals that are close to the original speech. The system is also able to generate speech with different emotions and speaking styles, demonstrating its potential for use in speech synthesis applica-tions. Overall, the proposed method provides a promising approach for generat-ing speech from video recordings of a face, which could have important applica-tions in areas such as the cinematography, the game development, virtual reality systems, rehabilitation, and medicine.
38. ALADEL ARIJ
The Moscow Institute of Physics and Technology (State University)
SAMDIT: Systematic study of adding memory to divided input in transformer to process long documents
n recent times processing long documents using a transformer has attracted the attention of research society. Different structure modifications and pre-training objectives were suggested. This paper introduces processing long documents using the segmentation method and added memory-segment representation. We have studied the usage of unified relative positional encoding for all tokens in a memory slot related to a chunk in the proposed transformer. This paper introduces a small overview of the positional encoding in the transformer and suggests positional encoding schemes for processing long documents, our method to process long documents using the combination of segmentation, unified relative positional encoding for slot tokens, masking technique, and additional memory slots related to segments. This method was tested on summarization tasks even though it can also be used for different tasks such as reading comprehension tasks and translation. The main results show on-par performance for using our method to process long documents with the baseline, and superior performance compared to SLED. Studying memory content reveals that unified memory initialization and unified positional encoding for memory slots led to unified content for all tokens in each slot, and active interaction between memory tokens and chunks. This interaction led to encoding noise at the end of encoding.
39. MIKHAIL KISELEV, ALEXANDER IVANITSKY, DMITRY IVANOV AND DENIS LARIONOV
1The Chuvash state university named after I. N. Ulyanov
2Lomonosov Moscow State University
3
A Spiking Neuron Synaptic Plasticity Model Optimized for Unsupervised Learning
Learning in spiking neural networks is implemented through synaptic plasticity. Diversity of various learning regimes assumes that different forms of synaptic plasticity may be most efficient for, for example, unsupervised and supervised learning. In the present paper, we formulate specific requirements to plasticity rules imposed by unsupervised learning problems and construct a novel plasticity model satisfying these requirements. This plasticity model serves as main logical component of the novel unsupervised learning algorithm called SCoBUL (Spike Correlation Based Unsupervised Learning).
40. * CHAPLINSKAIA N.V., BAZENKOV N.I.
V. A. Trapeznikov Institute of Control Sciences of Russian Academy of Sciences, Moscow
Training a spiking neural network with myelination and demyelination mechanisms
This paper considers a model of a spiking neural network with a plastic myelin sheath of axons, which is capable of regulating axonal delays in the propagation of spikes due to the processes of myelination and demyelination. The proposed mechanism synchronizes the reactions of the neuron postsynaptic potential to incoming spikes and contributes to the memorization of exposed patterns by neurons. A simulated spiking neural network was successfully trained to recognize handwritten digits from the USPS dataset.
41. DMITRY ANTONOV, SERGEY SUKHOV
Training of Spiking Neural Networks - classifiers by local rules
Spiking Neural Networks (SNN) are one of the promising candidates for solving the problem of reducing power consumption in the course of machine learning. Unfortunately, it is difficult to organize the backpropagation in Spiking Neural Networks. The present study demonstrates reinforcement learning method of SNN entirely based on local rules. The experiments were carried out in the Brian 2.0 environment on free available MNIST datasets.
42. R.B. RYBKA, D.S. VLASOV, A.I. MANZHUROV, A.V. SERENKO, A.G. SBOEV
1National Research Centre "Kurchatov Institute", Moscow
2The Moscow Institute of Physics and Technology (State University)
Spiking neural network with local plasticity and sparse connectivity for audio classification
The purpose of this work is to study the sparse connectivity method for a spiking neural network with training based on local plasticity. Various options for organizing connections in a three-layer spiked neural network are considered. In the proposed sparse connectivity method, connections between layers of neurons are formed with some probability within a limited area of neurons location. Testing of the resulting architectures with local plasticity was carried out on the audio classification.
43. A.YU. DOROGOV
Saint Petersburg Electrotechnical University "LETI"
Implementation of image memory in the class of fast neural networks
The application of backward-oriented pyramidal neural networks of fast learning for implementation of image memory elements is considered. Networks of the class under consideration are representable by linear operators, have a self-similar structure, and are a special case of the fast Fourier transform algorithm. Methods of topological construction of pyramidal networks are given. It is shown that the pyramidal memory network provides storage and recovery of images similar to storing numbers in random access computer memory. It is proved that pyramid networks belong to the category of networks with a deep degree of learning.
SESSION 5
Wednesday, October 25 14:00 – 17:15
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. USHAKOV VADIM
Neurobiology and neurobionics
44. * ZIMIN ILYA A., STASENKO SERGEY V.N.I. Lobachevsky State University of Nizhni Novgorod
Artificial neural network with dynamic synapse model
In this study, we present an innovative hybrid model for short-term memory that combines short-term synaptic plasticity, astrocytic modulation of synaptic transmission, and a convolutional neural network. By comparing it with the recurrent neural network, we find that the proposed model demonstrates superior efficiency in accurately modeling short-term memory.
45. SERGEY V. STASENKO, TATIANA A. LEVANOVA
N.I. Lobachevsky State University of Nizhni Novgorod
Mean-field model of brain rhythms controlled by glial cells
We propose a theoretical framework that describes how glial cells can regulate neuronal activity and contribute to the generation of brain rhythms. In the proposed model, glial cells modulate the excitability of neurons by releasing gliotransmitters. The model takes into account the collective behavior of a large population of neurons and describes how the interactions between neurons and glial cells can give rise to different patterns of synchronized activity, such as oscillations and waves.
46. ANDREY A. LEBEDEV, VICTOR B. KAZANTSEV, SERGEY V. STASENKO
N.I. Lobachevsky State University of Nizhni Novgorod
Study of the Influence of Synaptic Plasticity on the Formation of a Feature Space in the Inhibitory Layer of a Spiking Neural Network
This paper presents a novel model for image decoding using a spiking neural network for the recognition task. Our approach involves utilizing the average number of spikes in the inhibitory layer as a measure of the neural network's response to a sensory stimulus, which serves as a unique feature.
47. ALEXANDRA I. BULAVA, ZHANNA A. OSIPOVA AND YURI I. ALEXANDROV
Institute of Psychology of Russian Academy of Sciences, Moscow
Effect of Anxiety and Exploratory Activity on Fos Activated Neurons in the Deep and Superficial Layers of the Rat Cerebral Cortex
Although using neuroscience to develop artificial intelligence (AI) may guide neural network models toward human-like learning, at the moment artificial neural networks differ from the nervous system in many significant functional patterns. In order to succeed in creating AI with human-like cognitive abilities, neuro- and cognitive sciences should participate in AI research as a part of the joint research program. In this study we used c-Fos immunolabeling to identify the experience-dependent mismatch specific changes in cortical activity. To determine parameters of individuality and to assess the extent of relationship between different parameters, we analyzed the behavioral activity of rats in tests based on locomotion and anxiety levels. The results of this study demonstrated that experience-dependent mismatch induced cortical layer-specific changes in activity. Anxiety and exploratory activity were associated with selective changes in the number of Fos-activated neurons in the deep and superficial cortical layers, but were not associated with the total number of Fos-expressing cortical neurons in this area of the brain. We found a significant effect of anxiety and exploratory activity on learning rate. We argue that individual differences in learning can be predicted by the respective behavioral tests to measure exploratory and anxiety-related behavior.
48. ZHILYAKOVA LIUDMILA
V. A. Trapeznikov Institute of Control Sciences of Russian Academy of Sciences, Moscow
Modeling neuron-like agents with a network internal structure
In the paper, we describe the model of heterogeneous agents with an internal structure and a set of parameters. These agents can generate an endogenous activity in definite time intervals. During the activation phase, the agent emits a spe-cific mediator with an assigned color into a common medium. In case agents correspond to the biological neurons, this medium is an extracellular space, and mediators are neurotransmitters.
49. * ROGACHEV A. O., SYSOEVA O. V.
Application of the temporal response function to study the neurophysiological mechanisms of speech perception in ecologically valid conditions
We consider the application of the temporal response function, a method that allows to study the neurophysiological mechanisms of natural speech stimuli processing. In our study, children aged 6 to 8 listened to audio recordings of children's fairy tales, while recording an electroencephalogram. As a result, it was shown that the predictive coefficients of the temporal response function were positively related to the level of development of speech comprehension.
50. MYSIN I.E.
Institute of Theoretical and Experimental Biophysics of RAS, Pushchino
Adaptation of machine learning methodology for modeling neural networks of the brain
Neuroscience has accumulated a large amount of data on the functioning of the brain when processing information. However, it was not possible to generalize the volume of modern data in the form of mathematical models. We propose a new approach to building models of impulse neural networks. Our approach is based on combining two already known approaches: population models and automatic differentiation. It is shown that using gradient descent, it is possible to find network parameters that best describe experimental data.
51. I.A. SMIRNITSKAYA
Scientific Research Institute for System Analysis, Moscow
The role of pulvinar nucleus as a sinchronizer of cortical activity for visual target detection is caused by its role as superior colliculus–cortex intermediary
The brain appeared for the movement control. Evolution has turned its function into behavior control. In the course of evolution, the brain has evolved into an extremely complex hierarchically organized system. To create a simplified model that algorithmically describes its structure, it was proposed [8] to construct two parallel sequences: a sequence of models of brain structure representing the stages of its phylogeny, and a sequence of animal behavior models at the corresponding stages of phylogeny. As an illustration of the usefulness of such an approach for creating an operational model of a specific brain regions, the functions of the thalamic pulvinar nucleus in the control of visually guided behavior are discussed. It is concluded that the experimentally discovered role of the pulvinar as an initiator and synchronizer of parieto-frontal interactions is due to its main input signals from the superior colliculus and the pretectum
52. PROSKURA A.L., VECHKAPOVA S.O., RATUSHNYAK A.S.
Insulin as a glutamate receptors density modulator
The paper considers molecular rearrangements in the interactome of the pyramidal neuron of the hippocampal CA1 field dendritic spine under the hormone insulin control. A reconstruction of the insulin receptor signaling pathway was proposed that involved in the modulation of excitation synaptic neurotransmission . The protective effect of insulin and its contribution to the control of the increase in synaptic glutamate receptors in the hippocampus are discussed.
53. A.S. RATUSHNYAK, A.L. PROSKURA, S.O. VECHKAPOV, R.R. KHUSAINOV
Principles of modeling the origin and evolution of negentropic information systems
The problem of developing principles for modeling molecular complexes that have a function inherent in living things, the ability to preserve complex entropy, is considered. Based on the concept of maintaining stability when the feedback loop of the agent with the environment is closed, the structure, algorithm and code of the negen-tropic information system model have been developed. It was taken into account that the reaction of the system to external factors is determined not only by the presence of stimuli, but also by previous influences, i.e. memory and anticipatory reflection
54. VLADISLAV DOROFEEV
Scientific Research Institute for System Analysis, Moscow
Cognitive Functions of Cerebellum and Educational Neuroscience
With the development of methods of psychological and neurophysiological studies, the facts of the influence of disorders of the cerebellum on cognitive functions, anatomical and neurophysiological connections of the cerebellum with the prefrontal cortex responsible for cognitive functions were revealed, and theories of the involvement of the cerebellum in cognitive activity were constructed. The following is an overview of the current state of research in this area and demonstrates possible development of this area with educational neuroscience.
55. SYSOEV I. V., KORNILOV M. V., SYSOEVA M. V., EGOROV N. M.
Mathematical and radiophysical models of limbic epilepsy principal pattern generator
Limbic epilepsy is the most common form of the disease and is often accompanied by seizures. It is proposed that the source of pathological activity is located in the hippocampus of one of the hemispheres and is a tiny area or network. We have proposed a simple model in the form of a ring of connected pyramidal neurons, which makes it possible to implement a principal pattern generator with a tunable frequency in a numerical and natural radiophysical experiment.
56. N.G. BIBIKOV
N.N. Andreyev Acoustics Institute, Moscow
Neurons in the auditory cortex of an unanesthetized cat reproduce rapid changes in long-term signals
The reactions of single neurons in the auditory cortex of a cat to various complex signals (tones modulated by noise in the 0-50 Hz band, vocal and speech signals, sounds of potential preys) were studied under conditions close to natural (no anesthesia, no sound dampening). In contrast to the data obtained during any type of anesthesia, cortical neurons were capable of a long-term response to such signals. The ability to reproduce rapid temporal changes in the amplitude was retained throughout the long exposure.
POSTER SESSION 2
Wednesday, October 25 14:00 – 17:15
Lecture-hall НЛК 2 этаж
Chair: Prof. DOROGOV ALEXANDER YURYEVICH
Adaptive behavior and evolutionary modelling
57. ZAREMA B. SOKHOVA * AND VLADIMIR G. RED’KOScientific Research Institute for System Analysis, Moscow
Modeling of natural needs of autonomous agents
In this paper, a model of autonomous agents with basic biological needs is constructed and investigated. The population of agents functions in a cellular environment. Each agent has four needs: 1) safety, 2) food, 3) reproduction, and 2) research. The intensities of needs form the agent’s genotype and are expressed as values from the interval [0, 1]. The model was analyzed by computer simulation. It is shown that agents with motivations have advantages as compared with agents without motivations. The experiments also confirmed that the needs of food and reproduction are the most important for the survival of the population.
58. VLADIMIR B. KOTOV AND GALINA A. BESKHLEBNOVA
Scientific Research Institute for System Analysis, Moscow
Variable resistor under high-frequency signal
Variable resistors ("memristors") have the potential to become important building blocks for creating artificial brains. Many of the obstacles that stand in the way of creating practically useful variable resistors can be overcome by using suitable methods to control the state of the resistor. As such a method, state change under the action of a variable high-frequency signal is considered. An equation for the period-averaged oscillation state of a model resistor (a simple resistor element) is derived and investigated.
59. SAMER EL-KHATIB, YURI SKOBTSOV , SERGEY RODZIN
1Saint Petersburg State University of Aerospace Instrumentation
2Southern Federal University, Rostov-on-Don
Determination of optimal heuristic image segmentation coefficients for modified Ant Colony optimization method for complexly structured images
For image segmentation ant colony optimization algorithm is proposed.The heu-ristic image segmentation coefficients in Ant Colony optimization strongly affects the computational complexity of this method. The Hybrid Ant Colony Optimiza-tion (ACO) – k-means and optimal number of ants detection for complexly struc-tured images are considered in this paper. Complexly structured images are imag-es with a non-deterministic and nonlinear structure composed of many complex elements with important properties. Heuristic parameters affects to computational complexity of this method. Balansed parameters most preferred to segmentation method to reduce performance degradation. The modified Ant Colony Optimiza-tion (ACO) – k-means and optimal heuristic parameters investigation for com-plexly structured images are considered in this paper.
Artificial intelligence
60. CHE ZHANG, YAOWEN HUANG, ELIZAVETA K. SAKHAROVA, ANTON.I. KANEV, VALERY.I. TEREKHOVBauman Moscow State Technical University
GrowNet - Enhanced CurveNet for Tree Species Classification
Nowadays, remote sensing is widely used for large-scale forest surveys. The use of LiDAR (Light Detection and Ranging, LiDAR) made it possible to obtain detailed 3D point clouds of scanned areas, significantly in-creasing the efficiency of identification. The breakthrough in 3D object clas-sification has opened new opportunities for the practical application of deep learning methods to identify forest tree species, which is a key task for forest management. In this paper, we propose a CurveNet-based model more suita-ble for tree species classification - GrowNet. GrowNet uses a deterministic algorithm to generate sampling curves. The results of tree classification ex-periments show that the GrowNet model has the highest accuracy result of 86.5% compared to PointNet, PointNet++, and CurveNet models.
61. TIRSKIKH D.V., KONOVALOV V.P.
1
2The Moscow Institute of Physics and Technology (State University)
Zero-shot NER via Extractive Question Answering
Although the task of named entity recognition (NER) is usually solved as a sequence tagging problem via traditional supervised learning approaches which require the presence of a substantially sized annotated dataset, recent works aiming to utilize pretrained extractive question-answering (QA) models have shown significant few and zero-shot capabilities. This work aims to further investigate their applicability in zero-shot setting i.e without explicit fine-tuning.
62. KURYAN V.E.
Axioms of constructing a graph of the world model
The fundamental limitation of the "sequence in sequence" paradigm in the construction of artificial intelligence systems is noted. An approach to the construction of intelligent systems based on the use of a world model in the form of a graph, which is built automatically in the learning process, is described. The rules (axioms) of the automatic construction of the graph of the world model are formulated. It is shown that the resulting model has the property of explainability
63. MINAKOV G. A., AKHMADJONOV M. K. U., KUZNETSOV D. P.
1The Moscow Institute of Physics and Technology (State University)
2DeepPavlov
Dialogue Graphs: Enhancing Response Selection through Target Node Separation
This paper proposes a new method called Target Node Separation to address the problem of accurately performing the response selection task in dialogue systems. The proposed method enhances the performance of response selection tasks by refining the graph structure through improving edge ends. Authors compare the proposed method to other state-of-the-art methods on the MultiWoZ dataset and find that the proposed approach outperforms other graph-based methods and SBERTMap on recall metrics. Furthermore, the authors observed that increasing the number of clusters results in an improvement in the performance of the dialogue graph.
64. ANTON KOLONIN
Evolution of Efficient Symbolic Communication Codes
The paper explores how the human natural language structure can be seen as a product of evolution of inter-personal communication code, targeting maximization of such culture-agnostic and cross-lingual metrics such as anti- entropy, compression factor and cross-split F1 score. The exploration is done as part of a larger unsupervised language learning effort, the attempt is made to perform meta-learning in a space of hyper-parameters maximizing F1 score based on the “ground truth” language structure, by means of maximizing the metrics mentioned above. The paper presents preliminary results of cross-lingual word-level segmentation tokenization study for Russian, Chinese and English as well as subword segmentation or morpho-parsing study for English. It is found that language structure form the word-level segmentation or tokenization can be found as driven by all of these metrics, anti-entropy being more relevant to English and Russian while compression factor more specific for Chinese. The study for subword segmentation or morpho-parsing on English lexicon has revealed straight connection between the compression been found to be associated with compression factor, while, surprising, the same connection with anti- entropy has turned to be the inverse.
Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling
65. OLGA SOFRONOVA, DILYARA ZHARIKOVAThe Moscow Institute of Physics and Technology (State University)
Language Models Explain Recommendations based on Meta-Information
For recommender systems, the explanation of why the item was recommended to a user increases the reliability. In this work, we introduce a post-hoc method of explaining any recommender system's output with the use of LLMs and meta information about a recommended item and user's preferences. We try different models and introduce metrics for estimating the quality of generated explanations. The models are evaluated on three domains and then compared to analyze the ability for domain transfer.
66. IGOR M. ARTAMONOV, YANA N. ARTAMONOVA
1Moscow Aviation Institute (National Research University)
2NeuroCorpus
Analysis of text data reliability based on the audience reactions to the source
Classification between relevant and irrelevant data is one of the most important tasks in modern machine learning. This paper developes a semi-supervised classification methods that is based on partitially defined information for a text source. The method is based on the analysis of the author’s engagement into the subject area and the combination of his assessment by readers with quality of published texts. To achive this we used a combination of joint author and text analysis that allowed to significantly reduce workload for further data markup. The method uses a learning loop that balances class attribution probabilities for text data at the level of the most effective class separation for a given noise level in the data. It was found that different paths for estimating relevance probability for relevant and irrelevant records had to be used. Both likelihood and plausability approaches were used to achive acceptable level of classification. The method was developed in process of analysys of large number of texts from social networks. It showed high effectiveness on a large amount (> 70mln records, > 8mln unique authors) of confusingly similar text data.
67. JULIA EDELEVA, OLGA A. SMETANINA, A.A. MESHCHERYAKOVA, NADEZHDA KUSHINA
1Technische Universität Carolo-Wilhelmina zu Braunschweig
2Higher School of Economics, Nizhny Novgorod Branch
3
German-as-a-Foreign-Language (GFL) Learners’ Reli-ance on ‘Early’ and ‘Late’ Morphological Cues for Structural Analysis: The Role of Cognitive Load
We report the result of an eye-tracking study investigating cue reliance for structure building in LX learners of German at different proficiency levels. The eye movements of the participants were tracked on a three-character visual display while they were exposed to a subject or an object relative clause. After they listened to the stimulus sentence, the participants performed one of the two tasks. In one version, they had to immediately point at the intended character on the visual display. In an alternative version, they were subsequently presented with one of the three characters and had to validate or decline it as the intended/non-intended one. The latter task imposed a greater cognitive load on the participants since they had to memorize the selected character for the subsequent task. The numeric subject advantage around the relative clause offset was only evidenced in the less cognitively demanding task only. Additionally, the object relative clause condition evidenced more proportions of looks towards the embedded NP character. Thus, LX learners of German seem to emanate the processing pattern evidenced for monolingual German children and adult native speakers under a less demanding task. Our findings suggest that structural processing in morphologically rich languages is an interplay of ‘early’ and ‘late’ cues as well as the frequency of the structure itself. Simpler structures (e.g., subject relative clauses) can be computed provided there are sufficient cognitive resources while more complex structures remain challenging.
68. KRYLOV A.K.
Institute of Psychology of Russian Academy of Sciences, Moscow
Nonlogical reasoning and culture-dependence of logical task solving
The differences between human reasoning and mathematical logic are considered. It is shown the dependence of reasoning from culture type: western and eastern, and the according mentality. Logical reasoning as the process of logical tasks' solving depends on parenting in the childhood and on the environment, what makes it very different from the mathematical logical reasoning.
69. VLADIMIR N. SHATS
Independent investigator, St. Petersburg
Principle Splitting of Finite Set in Classification Problem
This paper deals with application of the concept feature similarity for objects of a finite set, which was proposed by the author, in the classification problem. On its basis, a method for solving the classification problem has been developed, which consists of separate stages of calculations: splitting ordered feature values into an equal number of intervals, calculating ordered pairs of object numbers and interval numbers, calculating granules containing ordered pairs of training sample objects for each class. Then, neglecting the difference in feature values within the intervals, we can calculate the frequency of granule objects for each feature using the simplest formulas of the event theory. These frequencies make it possible to calculate its frequency in each class by totality of the object's features, and then determine of the object class. Effectiveness of the method was demonstrated on 9 databases. Essentially, the method core is calculation of the surjective mapping of the given set onto the set of nested lists of object numbers. We consider the listed stages of mapping calculation as stages of implementing the principle partitioning of a data set into a set of nested lists of ordered pairs. This principle was developed when solving the problem of estimating closeness classes and clusters of a set calculated on its basis. The results of its solution for 10 combined databases indicate the promise of applying the principle.
70. BELOKOPYTOV ANTON SERGEEVICH, REDKOZUBOVA OLGA MIKHAILOVNA,
1National Research University of Electronic Technology
2National Research University "Higher School of Economics", Moscow
3Moscow State University of Psychology and Education
DEVELOPMENT OF AN ALGORITHM FOR DETECTING SLOW SPIKE-WAVE ACTIVITY IN ABSENCE EPILEPSY
This study focuses on the task of real-time detection of absence epilepsy seizures using electroencephalogram (EEG) data. A support vector machine (SVM) model was trained and tested based on the spectral characteristics of EEG windows containing seizures. The model achieved an accuracy of 0.97, sensitivity of 0.93, and specificity of 0.98. The algorithm can be integrated into a mobile application along with a wearable dry-electrode electroencephalograph (EEG) for real-time seizure detection.
Neurobiology and neurobionics
71. O.E. DICKPavlov Institute of Physiology of the Russian Academy of Sciences, St.Petersburg
Synchronization analysis of signals obtained from anesthetized rats during painful action
The task was set to reveal phase synchronization between such physiological rhythms as neuronal activity, fluctuations in blood pressure, and fluctuations in respiration in anesthetized rats. To solve this problem, the synchrosqueezed wavelet transform method was applied, which allows one to effectively calculate the instantaneous frequencies and phases of non-stationary signals. It was found that during painful action, the frequency of the neuronal activity variability or the frequency of the blood pressure variability can be adjusted at the respiratory rate with the subsequent occurrence of phase synchronization between these time series.
72. STASENKO SERGEY VICTOROVICH, KAZANTSEV VICTOR BORISOVICH
N.I. Lobachevsky State University of Nizhni Novgorod
Spiking neural network with tetrapartite synapse
In this paper we propose a model of a spiking neural network with a tetrapartite synapse, where astrocytic modulation induces neuron synchronization, and ECM activation leads to desynchronization.
73. VORONKOV GENNADY SERGEEVICH
Lomonosov Moscow State University
Is there an intracerebral screen for subjective visual images
The subjective visual image (SVI) of the real visual space (RVS) is not the result of the elements activity of the retina itself. It is generally accepted that the SVI is provided by the activity of the brain central visual structures. However, a neuron screen in the brain with an RVS image similar to that in the SVI was not detected. The paper attempts to identify in what objective form the representation of the exact coordinates of the RVS points is carried out in the brain, similar to that of the SVI.
74. LEONID I. BRUSILOVSKY, ANDREY S. BRYUKHOVETSKY, SERGEY P. KOZHIN
Studies of cognitive microwave radiations of the human brain as a new type of microwave encephalograms
Since 2016, a group of scientists and radio engineers has been proactively conducting research on the registration of microwave radiation of the human brain in the UHF/SHF microwave ranges. As a result, own electromagnetic radiation of the human brain was recorded in the range from 850 MHz to 5.0 GHz with signal power at the level of -130 dBm -100 dBm (10-16 - 10-13 W), which have cognitive features. The registration method was patented (patent RU 2708040 C2). Previous studies have determined the composition of the tools and the program-method of experiments. The main goal of the new study on July 12-13, 2021 was to confirm with representative experiments the cognitive nature of the recorded microwave bursts depending on the “light/dark” stimulus using an improved program-method with a multiple increase in the number of registrations per interval, as well as to search for new available tools, satisfying the patented registration method.
SESSION 6
Wednesday, October 25 17:45 – 19:30
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. RATUSHNYAK ALEXANDER SAVELYEVICH
Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling
75. ISABELLA G. SILKISInstitute of Higher Nervous Activity and Neurophysiology of RAS, Moscow
Mechanisms for contribution of modifiable inhibition to increasing the signal-to-noise ratio and contrasted representation of sensory stimuli in the neocortex
Proposed mechanism for long-term increasing signal-to-noise ratio in the various CNS structures is based on changes in postsynaptic processes evoked by simultaneous action of a neuromodulator on the same type of Gs or Gq/11 protein-coupled receptors located both on the main cell and input inhibitory interneuron. If initial excitation of the main cell is relatively strong, LTP is induced at its excitatory input sim-ultaneously with LTD at its inhibitory input. If excitation of the main cell is relatively weak, LTD is induced at its excitatory input simultaneously with LTP at its inhibitory input. Therefore, the reactions of the main cells will became stronger, and it will not respond to weak signals (which can be considered as a noise). Such effect is interpreted as in-creasing signal-to-noise ratio. Previously we explained why the sign (LTP or LTD) of the modifying action of dopamine on the efficacy of cortical inputs to spiny cells of the striatum (the input structure of the basal ganglia) depends on the initial strength of this excitatory input. Such character of dopamine-dependent modulations and subsequent activity reorganization in the cortico–basal ganglia–thalamocortical loops leads to a contrasting enhancement of neocortical representations of preferred sensory stimuli simultaneously with the weakening representations of other stimuli. This effect is facilitated by other neuromodulators due to an increasing signal-to-noise ratio on striatonigral spiny cells in combination with LTP of excitation of neurons in the neocortex, hippocampus, and thalamus projecting onto spiny cells. The proposed mechanisms are fundamentally different from generally accepted ones.
76. VITALY VERKHLYUTOV1, EVGENII BURLAKOV2 VICTOR VVEDENSKY3, KONSTANTIN GURTOVOY3, AND VLADIMIR USHAKOV4
1Institute of Higher Nervous Activity and Neurophysiology of RAS, Moscow
2V. A. Trapeznikov Institute of Control Sciences of Russian Academy of Sciences, Moscow
3National Research Centre "Kurchatov Institute", Moscow
4
Recognition of Spoken Words From MEG Data Using Covariance Patterns⋆
A speech recognition based on EEG and MEG data is an important step in the development of BCI and AI. Our approach for speech recognition relies on the evaluation of connections in the space of sensors with the identification of a pattern of MEG connectivity specific for a given segment of speech. We tested our method on 7 subjects. In all cases, the processing pipeline was quite reliable and worked either with- out recognition errors or with a small number of errors. After “training”, the algorithm is able to recognise a fragment of oral speech with a single presentation. MEG bandpass filtering showed that the quality of recognition is higher when using the gamma frequency range in comparison with the low-frequency range of the analysed signal.
77. * LEKHNITSKAYA POLINA ALEKSANDROVNA
Kazan (Volga region) Federal University
Model of non-visual eye-movements during performing cognitive tasks in short-term memory
In the blank screen paradigm participants solved cognitive tasks. Cur-rent fixation duration and saccade peak velocity differ in learning and per-forming some kinds of tasks. Retrieving information from memory and fol-lowing mental processing depends on the characteristics of input task and its difficulty. The best fixation accuracy performance was in the partici-pants who named only two levels of task difficulty, the best saccade accuracy performance was in Random Forest Classifier for saccades.
78. * KONONOV R.A., MASLENNIKOV O.V.
The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
Population dynamics in recurrent neural networks: reinforcement learning for cognitive neuroscience tasks
The paper presents a recurrent neural network that is trained by reinforcement learning methods to perform a task inspired by the context-dependent decision making experiment in the field of cognitive neuroscience. The dynamic mechanisms of the network execution of the objective function are studied. Functional clusters of neurons have been identified that are specialized in performing certain trial stages and distinguishing different features of input and output signals.
79. * ANASTASIA BRUEVA, JULIA EDELEVA
1N.I. Lobachevsky State University of Nizhni Novgorod
2Technische Universität Carolo-Wilhelmina zu Braunschweig
Negotiating cognitive load through cognitive skills, foreign language proficiency and the state of anxiety in L2 oral performance
Many foreign language learners experience a state of anxiety when taking foreign language exams or communicating in a second language. The state of anxiety might be conditioned by an increased cognitive load during L2 comprehension and production which has been considered in a number of research papers [4,10,11,18]. At the same time, cognitive load is closely re-lated to cognitive skills and, therefore, can be mediated both by external factors such as task complexity and internal factors such as working memory capacity. In the current study, we investigate whether the state of anxiety during L2 oral performance can be relieved as a function of working memory capacity. We elicited spontaneous L2 and L1 oral production data of 16 Russian-speaking participants. Adapted methodology originally proposed by Sineokova [23] was employed to annotate linguistic features of speech produced in a state of anxiety and in a neutral state. Selected disfluency types were identified that are associated with either working memory capacity or with the state of anxiety. Our results reveal that the interrelation between situational and personal anxiety, working memory capacity and the level of L2 proficiency is not statistically significant. Yet, general numeric trends could be observed. The findings have significant implications for the development of automated models of cognitive load estimation. The taxonomy of linguistic features can be used in language teaching and trainings on anxiety reduction through cognitive skill development. Further, the data might be useful to improve machine learning algorithms of affective state estimation based on oral speech data.
80. * MARKOVA G.M., BARTSEV S.I.
1Siberian Federal University, Krasnoyarsk
2
Does a recurrent neural network form recognizable representations of a fixed event series?
We study the identification of fixed sequences of events, which is processed by a model object – a simple recurrent neural network, by its neural activity using the neural network-based decoding method. Successful identification indicates that the recurrent neural network forms recognizable representations of the event series.
81. POLEVAIA S.A.,PARIN S.B,KOVALCHUK A.V.,FEDOTCHEV A.I.,EREMIN E.V.,PERMYAKOV S.A.
1N.I. Lobachevsky State University of Nizhni Novgorod
2The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
3Institute of Cell Biophysics, Pushchino, Moscow region
4Nizhny Novgorod State Medical Academy
5
Language of rhythmograms for displaying the functional states of a human
With the help of a convolutional neural network, a number of functional states were identified according to cardiorhythmograms obtained using the technology of event-related heart rate telemetry.
SESSION 7
Thursday, October 26 10:00 – 13:00
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. DOLENKO SERGEY
Applied neural systems
82. IGOR ISAEV, IVAN OBORNEV, EUGENY OBORNEV, EUGENY RODIONOV, MIKHAIL SHIMELEVICH AND SERGEY DOLENKO1Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
2Russian State Geological Prospecting University n. a. Sergo Ordzhonikidzе, Moscow
The Use of A priori Information in the Neural Network Solution of the Inverse Problem of Exploration Geophysics
This study is devoted to solving inverse problems of exploration geophysics, which consist in reconstructing the spatial distribution of the properties of the medium in the thickness of the earth from the geophysical fields measured on its surface. We consider the methods of gravimetry, magnetometry, and magnetotelluric sounding, as well as their integration, i.e. simultaneous use of data from several geophysical methods to solve the inverse problem. To implement such integration, in our previous studies we have proposed a parameterization scheme that describes a layered geophysical model with fixed layer properties, in which the determined parameters were the positions of the boundaries between the layers. In the present study, this parameterization scheme is complicated so that the properties of the layers vary from pattern to pattern in the data set. To improve the quality of neural network solution of the described inverse problem, we consider an approach based on the use of a priori information about the physical properties of the layers, in which this information is used directly as additional input features for the neural network.
83. * KARIMOV E.Z., SHIROKII V.R., BARINOV O.G., MYAGKOVA I.N.
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Domain Adaptation of Spacecraft Data in Neural Network Prediction of Geomagnetic Dst Index
This study focuses on improving the neural network prediction of the geomagnetic indexes, in particular Dst-index, in a scenario, where input data is collected by two spacecraft (SC) with different data availability. One of the SC is approaching the end of its operational lifespan, while the other one lacks sufficient data history for constructing a high-quality neural network prediction. To effectively perform the transition between the two SC data, domain adaptation methods are needed. The study evaluates and compares various data translation techniques and optimizes the parameters for each translated feature to minimize domain discrepancies. The findings highlight the enhancement in the forecast, when employing domain adaptation methods and selecting relevant features, surpassing the results obtained using untranslated data.
84. * ROMAN VLADIMIROV, VLADIMIR SHIROKIY, IRINA MYAGKOVA
1
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Investigation of the Importance of Input Features in the Problem of Predicting Geomagnetic Disturbances
In this paper, we study an algorithm for obtaining the most efficient model for predicting the amplitude of the geomagnetic Dst index, based on lowering the input data dimension by gradually discarding input features. This task is relevant, since the selection of significant input data is necessary for the effective use of machine learning methods. The study was carried out on the basis of the following machine learning methods: artificial neural network of the multilayer perceptron type, gradient boosting, linear regression. Comparison of the effectiveness of the listed methods is carried out.
85. * GUSKOV A.A., ISAEV I.V., BURIKOV S.A., DOLENKO T.A., LAPTINSKIY K.A., DOLENKO S.A.
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Integration of data from various physical methods in solving inverse problems of spectroscopy of solutions by machine learning methods
This article presents the results of solving the inverse problem of determining the concentrations of heavy metal ions of multicomponent solutions by Raman spectra, infrared spectra and absorption spectra using integration of optical spectroscopy methods. It is shown that the joint use of data from various physical methods make it possible to reduce the error of spectroscopic determination of concentrations of heavy metal ions in solutions. If the integrated methods differ significantly by their accuracy, then their integration is not effective. These effects are observed using various machine learning methods: random forest, gradient boosting and artificial neural networks – multilayer perceptrons. A series of experiments with solutions based on river water are also performed to estimate the variability of the fluorescence of natural waters in Moscow. A significant increase in the error level relative to solutions prepared in distilled water is observed, indicating the need to develop new methods to improve the quality of the solution of the investigated problem for the diagnostics of real river waters.
86. KHRYASHCHEV VLADIMIR
P.G. Demidov Yaroslavl State University
Neural network algorithms for medical decision support system in endoscopy investigation of stomach
At the present stage, medical decision support systems are widely used in endoscopic examinations. In the present study, the problem of detecting gastric pathologies on endoscopic video data is considered. An analysis of the efficiency of using the architectures of convolutional neural networks SSD and EfficientDet in relation to this problem has been carried out. 54 videos of stomach studies were used to train and test deep machine learning algorithms.
87. * CHUGREEVA GALINA N., SARMANOVA OLGA E., LAPTINSKIY KIRILL A., BURIKOV SERGEY A., DOLENKO TATIANA A.
Lomonosov Moscow State University
Application of convolutional neural networks for creation of photoluminescent carbon nanosensor for heavy metals detection
The paper presents results of the use of convolutional neural networks for the development of a multimodal photoluminescent nanosensor based on carbon dots (CD) for simultaneous measurement of the number of parameters of multicomponent liquid media. It is shown that using 2D convolutional neural networks allows to determine the concentrations of anions and pH value of aqueous solutions with errors that satisfy the needs of monitoring the composition of technological and industrial waters.
88. I G SHELOMENTSEVA
Realization of Super-Resolution using bicubic interpolation and an efficient subpixel model for preprocessing low spatial resolution microscopic images of sputum
Medical imaging explores methods and models for analyzing medical image data, however, the low resolution of images obtained using equipment with small lenses and a short focal length may limit the implementation of medical data recognition. A variety of models and methods of super resolution implement the preprocessing of low spatial resolution images in medical imaging. The paper in-vestigates the problem of preprocessing microscopic images of sputum contain-ing small-sized objects of interest using super-resolution methods of bicubic in-terpolation and a model of an effective sub-pixel convolutional neural network. The performance of the selected models and methods is evaluated using the PSNR criterion. The obtained results show that both approaches can be used for the problem of super resolution of microscopic images of sputum containing small objects of interest.
89. * SERGEY LINOK, DMITRY YUDIN
The Moscow Institute of Physics and Technology (State University)
Influence of Neural Network Receptive Field on Monocular Depth and Ego-Motion Estimation
We present an analysis of a self-supervised learning approach for monocular depth and ego-motion estimation. Unlike other existing works, our proposed approach called ERF-SfMLearner examines the influence of the deep neural network receptive field on the performance of depth and ego-motion estimation. We demonstrate on the KITTI dataset that increasing the receptive field leads to better metrics and lower errors both in terms of depth and ego-motion estimation.
90. * SHLAPAK N.P.
Research of the possibility of using a pre-tained autoencoder for principal component analysis
The possibility of using a neural network model to extract the principal components from a data array is considered using the example of the distribution of the energy release field in the core of the VVER-1000 reactor. A comparison with deterministic algorithms was made, an algorithm for model preparation was proposed, developed and tested
91. * ILYA A. GRISHIN, TIMOFEY Y. KRUTOV, ANTON I. KANEV AND VALERI I. TEREKHOV
Bauman Moscow State Technical University
Individual Tree Segmentation Quality Evaluation using Deep Learning Models LiDAR Based
The study of the forest structure makes it possible to solve many important problems of forest inventory. LiDAR scanning is one of the most widely used methods for obtaining information about a forest area today. To calculate the structural parameters of plantations, a reliable segmentation of the initial data is required, the quality of segmentation can be difficult to assess in conditions of large volumes of forest areas. For this purpose, in this work, a system of correctness and quality of segmentation was developed using deep learning models. Segmentation was carried out on a forest area with a high planting density, using a phased segmentation of layers using the DBSCAN method with preliminary detection of planting coordinates and partitioning the plot using a Voronoi diagram. The correctness model was trained and tested on the extracted data of individual trees on the PointNet ++ and CurveNet neural networks, and good model accuracies were obtained in 89% and 88%, respectively, and are proposed to use the quality assessment of clustering methods, as well as improve the quality of LiDAR data segmentation on separate point clouds of forest plantations by detecting frequently occurring segmentation defects.
92. * VYACHESLAV RAZIN, ALEXANDER KRASNOV, DENIS KARCHKOV, VIKTOR MOSKALENKO, DENIS RODIONOV, NIKOLAI ZOLOTYKH, SMIRNOV LEV AND OSIPOV GRIGORY
N.I. Lobachevsky State University of Nizhni Novgorod
Solving the problem of diagnosing a disease by ECG on the PTB-XL dataset using deep learning
Diagnosis by electrocardiogram (ECG) is an extremely urgent and important task, the quality, timeliness and speed of which people's lives and health depend on. To date, a large number of researchers treat neural networks as a panacea, hoping that any task can be solved quickly and without problems. Often, this approach does not lead to the best results. The article explores the use of deep learning as a universal tool for solving the problem in determining pathological ECG signals with markers of myocardial infarction, hypertrophy, conduction disturbances, and changes in ST segment morphology. During the experiments, the positive impact of using thresholding and replacements to increase the predictive ability of the network, the use of various ensembles on trained deep learning models was established. The addition of artificial models also improves the classifying ability of ensembles. Returning a random number in the absence of a single mode also makes it possible to increase the accuracy of the ensemble.
93. * MARAT SAIBODALOV, IAKOV KARANDASHEV
Scientific Research Institute for System Analysis, Moscow
Application of self-supervised learning algorithms for anomalies detection of in X-rays images
In this paper, we consider the problem of finding foreign objects in X-rays obtained from personal screening scanners. Such scanners are used at facilities requiring increased security control. The available data have a number of problems, which are described and addressed in the text. In this paper, only self-supervised anomaly detection algorithms that are trained on data that does not contain anomalies will be considered. Such models, unlike supervised learning algorithms, do not require a large amount of labeled data for training.
SESSION 8
Friday, October 27 11:00 – 13:00
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. BAKHSHIEV ALEKSANDR VALERYEVICH
Applied neural systems
94. FOMIN I.S., KORSAKOV A.M., IVANOVA V.V., BAKHSHIEV A.V.1Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
2Saint Petersburg Electrotechnical University "LETI"
3Peter the Great St. Petersburg Polytechnic University
Investigation of a spike segment neuron in the offline multi-object tracking task with embeddings constructed by a convolutional network
The problem of multi-object offline object tracking is considered. A combination of a convolutional neural network for translating images into an embedding space and a spike neural network for classifying image vectors in the embedding space is used. An accuracy of up to 48% is achieved when using half of the trajectory for training.
95. E.A. ILLARIONOV
Lomonosov Moscow State University
Adaptive neural network model for handwriting digitization
Automatic handwriting digitization is still a difficult task, including for machine learning models. One factor that negatively affects the accuracy of the model is the high variability in the handwriting of different authors. In the standard approach, machine learning models are trained on a pre-trained sample. When working with handwritten text, it is usually difficult to guarantee that the training sample covers the variety of handwritings that the model has to work with. In practice, this leads to the fact that texts with new handwriting, unlike the examples that were in the sample, are recognized significantly worse. In our paper we propose an idea how to implement an adaptive neural network model, which will automatically adapt to new handwriting styles and other features of the document. The proposed idea, of course, is not universal, but in some tasks it can significantly improve the quality of text digitization. In this work, this model is used to digitize multi-page manuscript catalogs with tabular data on solar observations.
96. YURY V. TIUMENTSEV, ROMAN A. TSHAI
Moscow Aviation Institute (National Research University)
SNAC Approach to Aircraft Motion Control
Dynamic programming, as a basis for reinforcement learning, is a well-known method for synthesizing control laws for dynamical systems. However, this approach suffers from the so-called ``curse of dimensionality,'' due to which it is of limited suitability for solving real-world problems, in particular, aircraft flight control problems. One way to overcome this drawback is to use Approximate Dynamic Programming (ADP), which combines reinforcement learning and feed-forward neural networks. As one of ADP variants ACD (Adaptive Critic Design) approach has been introduced and actively continues to develop. It is based on the concept of adaptive critic and exists in a large number of varieties. One of these varieties is called SNAC (Single Network Adaptive Critic). The specific feature of the SNAC approach is the use of a single neural network of a critic to be trained, which reduces the consumption of resources for forming the required control law. In this case the absence of a actor network as a part of SNAC system is compensated by using a special kind of optimization algorithm. This paper analyzes the essence of the SNAC approach, as well as the features of its implementation as applied to the control of a nonlinear dynamical system under uncertainty conditions. The capabilities of this approach are demonstrated on the example of applied problem, in which the control law of longitudinal angular motion of a passenger aircraft is synthesized. The results allow us to evaluate the effectiveness of the SNAC approach, as well as to identify its elements that require further research and development.
97. I. A. SOLOVIEV, V. V. KLINSHOV
The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
Stability thresholds of attractors of the Hopfield network
Purpose of the work is the detailed study of the attractors of the Hopfield network and their basins of attraction depending on the parameters of the system: the size of the network and the number of stored images. To characterize the basins of attraction so-called stability threshold method has been used, i.e., the method of searching minimum distance from an attractor to the boundary of its basin of attraction. For useful attractors, this value corresponds to the minimum distortion of the stored image, after which the system is unable to recognize it. It is shown that the dependence of the average stability threshold of useful attractors on the number of stored images can be nonmonotonic, due to which the stability of the network can improve when new images are memorized. An analysis of the stability thresholds allowed to estimate the maximum number of images that the network can store without fatal errors.
98. BOLOTNIKOVA A.A., SHLAPAK N.P.
1Obninsk Institute for Nuclear Power Engineering
2
Application of neural networks in prediction of changes in the isotope composition of fuel assembly of the VVER-1000 reactor
The possibility of using neural networks in solving the problem of predicting the isotopic composition of nuclear fuel is considered. An algorithm based on a recurrent neural network has been proposed, developed and tested, which is capable of predicting the isotopic composition of fuel assemblies at the requested time point using such parameters of fresh fuel as enrichment and mass content of gadolinium oxide in fuel rods. This algorithm can be used to perform quick preliminary calculations.
99. KORSAKOV ANTON, ISAKOV TIM, BAKHSHIEV ALEKSANDR
1Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
2Peter the Great St. Petersburg Polytechnic University
The Strategy of Incremental Learning on a Compartmental Spiking Neuron Model
The article presents a method for implementing incremental learning on a compartmental spiking neuron model. The training of one neuron with the possibility of forming new classes was chosen as an incremental learning scenario. During the training, only a new example was used, without knowledge of the entire previous training sample. The results of experiments on the Iris dataset are presented, demonstrating the applicability of the cho-sen strategy for incremental learning on a compartmental spiking neuron model.
100. EKATERINA A. ENGEL, NIKITA E. ENGE
Katanov Khakass State University, Abakan
An intelligent day ahead solar plant's power forecasting system
The power production of a photovoltaic system has complex nonlinear dynamic with uncertainties since solar radiation and temperature fluctuate. Thereby, it is complicated to approximate this complex dynamic by conventional algorithms while machine learning algorithms provide the required forecast’s performance. We solved the intelligent day ahead solar plant’s power forecasting task based on the modified neural net with a developed fuzzy attention mechanism. In contrast with existed fuzzy attention mechanisms which use classical membership func-tion we consider an attention mechanism’s context vector as fuzzy measures. The developed fuzzy attention mechanism selects from all signals which provided by classical attention mechanism only important signal based on fuzzy measures and the fuzzy integral. The developed fuzzy attention mechanism selects from all sig-nals which provided by classical attention mechanism only important signal based on fuzzy measures and the fuzzy integral. The comparison simulations study re-sults of the intelligent day ahead solar plant’s power forecasting system based on the created modified neural net with a fuzzy attention mechanism reveal its ad-vantages and competitive performance, as compared to a classical RNN. The modified neural net with a fuzzy attention mechanism has competitive perfor-mance as compared to a modified fuzzy neuronet and RNN for day ahead fore-casting of the hourly solar plants’ power.
101. PETR KUDEROV, EVGENII DZHIVELIKIAN, ALEKSANDR I. PANOV
1The Moscow Institute of Physics and Technology (State University)
2Federal Research Center "Informatics and Control" RAS, Moscow
Attractor Properties of Spatiotemporal Memory in Effective Sequence Processing Task
In this paper, we propose a biologically plausible model of spatiotemporal memory with an attractor module and study its ability to efficiently encode sequences, extract and reuse repetitive patterns. The results of experiments on synthetic and textual data, as well as on data from DVS cameras demonstrate a qualitative improvement in the properties of the model when using the attractor module.
SESSION 9
Friday, October 27 14:15 – 16:00
Lecture-hall НЛК 2 этаж, Конференц-зал
Chair: Prof. SAMSONOVICH ALEXEI VLADIMIR
Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling
102. STANKEVICH LEV ALEXANDROVICHPeter the Great St. Petersburg Polytechnic University
Cognitive neuro-fuzzy control systems
The work is devoted to the problems of developing cognitive neuro-fuzzy control systems. Systems are considered in which the cognitive functions of predicting and classifying the states of the environment for the control object are implemented. It is shown that the existing classifiers can provide an accuracy of 60-80% for 4 classes of the states. A new type of classifier based on a neuro-fuzzy network has been proposed, which showed an accuracy of the state classification at least 80%. Examples of using the classifier into cognitive control system of robots are given
103. VICTOR VVEDENSKY, VITALY VERKHLYUTOV, KONSTANTIN GURTOVOY
1National Research Centre "Kurchatov Institute", Moscow
2Institute of Higher Nervous Activity and Neurophysiology of RAS, Moscow
Permanent sharp switches in brain waves during spoken word recognition
. We measured magnetic signals of the human brain during recognition of spoken words and permanently observed clear jumps in the rate of the signal amplitude changes. The curves can be always represented as a concatenation of piecewise linear time segments with abrupt changes of slope. Duration of these time segments is highly variable, though it follows common dependence for all segments of the whole record. The brain signals during execution of a cognitive task resembled relay system switching permanently when performing computa-tions. We intend to reveal links between remote areas in the brain which syn-chronize moments of abrupt switching.
104. KOZUNOV VLADIMIR
Neurocognitive Research Center (MEG Center), Moscow
Spatial resolution for representational similarity analysis using magnetoencephalography
Representational similarity analysis provides ample opportunities for interpreting data obtained using various techniques for measuring brain activity. However, its application to correlate representations devoid of spatial or temporal dimension may be accompanied by the loss of information potentially carried by individual neuroimaging modalities. A method is proposed for localization in the brain of activity that corresponds to representations derived exclusively from magnetoencephalography data.
105. MONAHHOVA ELIANA, KLUCHAREV VASILY ANDREEVICH, MOROZOVA ALEXANDRA NIKOLAEVNA, BREDIKHIN DIMITRY OLEGOVICH, SHESTAKOVA ANNA NIKOLAEVNA, MOISEEVA VICTORIA VLADIMIROVNA
National Research University "Higher School of Economics", Moscow
Neurocogntive processing of attitude-consistent and attitude-inconsistent deepfakes: N400 study
The project examines behavioral and electrophysiological brain responses to auditory deepfakes taking into account congruence or incongruence of internal attitudes and the degree of analytical thinking, need for cognition and conformity of participants. We found that the level of trust is significantly influenced by speaker, interaction between attitudes and speaker and conformity factors and the higher negativity peaking was observed for pro-vaccination group in mismatch to internal attitudes and public opinion of speaker, similar to the N600-effect.
106. KOZIN A.V., GERASIMOV A.K., PAVLOV A.V., BAKAEV M.A.
Novosibirsk State Technical University
The development of the device for representing photostimuli in brain-computer interfaces based on the SSVEP paradigm
This article addresses the problem and discusses the peculiarities of presenting photostimuli in brain-computer interfaces based on the SSVEP paradigm. A universal device for precise adjustment of stimulation parameters has been proposed and developed. The results of the experiments underline the importance of taking into account individual reactions of users to the presented photostimuli in designing of brain-computer interfaces based on SSVEP potentials.
107. EKIZYAN A. KH., SHAPOSHNIKOV P.D., KOSTULIN D.V., SHAPOSHNIKOV D.G., KIROY V.N.
1Scientific Research Center of Neurotechnologies, Southern Federal University, Rostov-on-Don
2Southern Federal University, Rostov-on-Don
Real-time movement-related EEG phenomena detection for portable BCI devices. Neural network approach
In recent years, interest for brain computer interfaces (BCI) and their potential applications has been grown. However, despite their potential benefits, there are still many challenges which should be solved before BCIs can be widely used outside of laboratory conditions. One of the key issues is the real-time discrimination of movement-related EEG phenomena, which is essential for the use of portable EEG devices in everyday life. In this study different machine learning approaches with preliminary statistical and spectral feature extraction were compared in classification of movement-related artifacts. Dataset in this research was obtained from experiment with portable EEG of our development. Tested methods demonstrated high accuracy up to 80 percent in 7-classes discrimination task.
108. VLADIMIR KLINSHOV, ANDREY KOVALCHUK
The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
Rate chaos and memory lifetime in spiking neural networks
Frequency chaos is a collective state of a neural network characterized by slow, irregular firing frequencies of individual neurons. We study a sparsely connected network of spiking neurons that detects three different occurrences of frequency chaos. All cases lead to the fact that the collective dynamics becomes apparent.
